Why does AI-generated creative writing feel worse?

A recent dev.to post reframes the debate over declining AI writing quality, arguing the culprit isn't speed optimizations like speculative decoding, but a deeper training flaw called mode collapse.

Where it happened

A detailed technical breakdown by software developer Dann Waneri on the platform dev.to, published in early July 2026, serves as the anchor for this debate. The post, titled "Why AI Still Can't Write Well and Which Half of That Problem Is Actually Yours," responds to a theory Waneri describes as having gone "semi-viral" in the preceding weeks. The discussion is less a single thread and more a recurring, technically-plausible explanation for a widely-felt decline in the creative output of major language models.

Side A: Inference optimizations are the culprit

This argument posits that the degradation in AI writing quality is a direct result of inference-time optimizations. To serve millions of users quickly, AI labs use techniques like speculative decoding. In this process, a smaller, faster model generates draft text which a larger, more powerful model then verifies. The intuitive conclusion is that this shortcut must come at a cost. A system optimized for speed over methodical, token-by-token generation is likely to produce less nuanced, less creative, and more generic output. The logic follows a familiar engineering trade-off: you can have it fast, cheap, or good, but not all three. Proponents of this view suggest that the push for faster response times has quietly compromised the very quality that made the models impressive in the first place.

Side B: The problem is mode collapse from RLHF

This side, articulated by Waneri, argues the focus on inference speed is a red herring. The core claim is that speculative decoding is mathematically lossless by design. The final output is identical to what the large model would have produced on its own; the technique only changes the speed of arrival, not the content. The real problem is a phenomenon called mode collapse, which stems from the training process itself, specifically Reinforcement Learning from Human Feedback (RLHF). During RLHF, human raters reward the model for producing helpful and harmless responses. Over time, this nudges the model away from a wide range of possible outputs toward a narrow band of safe, agreeable, and stylistically average text. Waneri cites a paper by Kirk et al. which measured this effect, finding that RLHF-trained models show significantly lower output diversity. The issue is not the delivery mechanism, but the training incentive itself. As Waneri puts it, "rater preference, in bulk, rewards smooth and safe over sharp and specific."

What's underneath

The debate highlights a fundamental challenge in AI development: the difficulty of creating a scalable, automated signal for subjective quality. Writing has no direct equivalent to a program that compiles or a math problem that is verifiably correct. In the absence of an objective quality metric, training relies on human preference, which, when aggregated, tends to sand down interesting edges and favor inoffensive mediocrity. The search for a technical scapegoat like speculative decoding is a way of avoiding the much harder problem. The underlying tension is not between speed and quality at inference, but between scalability and genuine creativity during training.

The investor read

This debate signals a potential ceiling for the 'scale and RLHF' paradigm that has dominated foundation models. If the primary training method for making models 'helpful' also makes them creatively generic, a significant market opportunity exists for companies that can solve this quality problem. This could involve developing novel training methods that don't rely on aggregate human preference, or creating highly-specialized, fine-tuned models for creative and technical domains where 'average' is a synonym for 'useless.' The value may shift from raw model access to the sophisticated post-training or alternative training techniques that can escape mode collapse and deliver genuinely differentiated output.

Pull quote: “Rater preference, in bulk, rewards smooth and safe over sharp and specific.”